Forecasting the term structure of option implied volatility: The power of an adaptive method

We model the term structure of implied volatility (TSIV) with an adaptive approach to improve predictability, which treats dynamic time series models of globally time-varying but locally constant parameters and uses a data-driven procedure to find the local optimal interval. We choose two specifications of the adaptive models: a simple local AR (LAR) model for a univariate implied volatility series and an adaptive dynamic Nelson–Siegel (ADNS) model of three factors, each based on an LAR, to model the cross-section of the TSIV simultaneously with parsimony. Both LAR and ADNS models uniformly outperform more than a dozen alternative models with significance across maturities for 1–20 day forecast horizons. Measured by RMSE and MAE, the forecast errors of the random walk model can be reduced by between 20% and 60% for the 5 to 20 days ahead forecast. In terms of prediction accuracy of future directional changes, the adaptive models achieve an accuracy range of 60%–90%, which strictly dominates the range of 30%–59% of the alternative models.